Interpretable Human-Label-Free Deep Learning for Real-Bogus Classification in Time-Domain Astronomy

Forum topic · 小凯 · 2026-07-08

Summary

A new paper (arXiv:2607.05393) by Raphaël Bonnet-Guerrini, Bruno Sanchez, Dominique Fouchez, and colleagues presents a deep learning framework for real-bogus classification in time-domain astronomy that requires no human-labeled data. Real-bogus classification, which separates genuine transient candidates from artifacts, is a critical step in automated discovery pipelines, yet reliable labels are expensive and community labels tend to be noisy and survey-dependent. The authors train on injected transients combined with artifact-dominated survey data, showing robustness under strong label contamination and delivering calibrated uncertainty quantification. The approach uses a dual-network model with asymmetric co-teaching to handle class-dependent label noise, analyzes learned representations via latent space visualization, and introduces a low-cost hybrid uncertainty quantification strategy that exploits the dual-network setup to improve calibration. The results demonstrate that injection-driven weak supervision enables scalable and consistent real-bogus classification without manual annotation.

Overview

Field: AI / Time-domain Astronomy
Authors: Raphaël Bonnet-Guerrini, Bruno Sanchez, Dominique Fouchez, Benjamin Racine, Maya Guy, Mariam Sabalbal, Manal Yassine, Vincenzo Piuri
Published: 2026-07-06
arXiv: 2607.05393

Abstract (translated)

Time-domain surveys produce vast numbers of transient candidates, and real-bogus classification is a critical step in automated discovery pipelines. Reliable labels are costly, while community labels are often noisy and survey-dependent. This work develops a real-bogus classification framework that does not require human-labeled data, using injected transients and artifact-dominated survey data for training. The framework remains robust under strong label contamination and provides calibrated uncertainty quantification.

Key points

Human-label-free training: The model is trained purely on injected transients plus artifact-dominated survey data, avoiding expensive manual annotation.
Asymmetric co-teaching: A dual-network architecture handles class-dependent label noise at different noise levels.
Interpretability: Learned representations are analyzed through latent space visualization.
Hybrid uncertainty quantification (UQ): A low-cost strategy leverages the dual-network setup to improve calibration of uncertainties.
Result: Injection-driven weak supervision enables scalable, consistent real-bogus classification without human labels.

---

*Source: arXiv:2607.05393. Auto-collected 2026-07-06.*